Autonomous Surveillance Tolerant to Interference

Autonomous recognition of human activities from video streams is an important aspect of surveillance. A key challenge is to learn an appropriate representation or model of each activity. This paper presents a novel solution for recognizing a set of predefined actions in video streams of variable durations, even in the presence of interference, such as noise and gaps caused by occlusions or intermittent data loss. The most significant contribution of this solution is learning the number of states required to represent an action, in a short period of time, without exhaustive testing of all state spaces. It works by using Surprise-Based Learning (SBL) to reason on data (object tracks) provided by a vision module. SBL autonomously learns a set of rules which capture the essential information required to disambiguate each action. These rules are then grouped together to form states and a corresponding Markov chain which can detect actions with varying time duration. Several experiments on the publicly available visint.org video corpora have yielded favorable results.